Thanks to visit codestin.com
Credit goes to github.com

Skip to content

训练到大概 9000 次迭代的时候 loss 变得巨大 #4

@Monalissaa

Description

@Monalissaa

再次感谢您的工作!
我按照您的默认设置(SAIDT数据集是按照论文里的训练测试方式进行划分的)进行了训练,如下所示

parser.add_argument("--save", default='./log', type=str, help="Save path")
parser.add_argument("--resume", default="", type=str, help="Resume path (default: none)")
parser.add_argument("--scale_factor", type=int, default=4, help="scale")
parser.add_argument("--input_num", type=int, default=7, help="input frame number")
parser.add_argument("--train_dataset_dir", default='./data/train/SAITD', type=str, help="train_dataset")
parser.add_argument("--val_dataset_dir", default='./data/test/SAITD', type=str, help="train_dataset")
parser.add_argument("--batch_size", type=int, default=2, help="Training batch size")
parser.add_argument('--patch_size', type=int, default=64)
parser.add_argument('--n_iters', type=int, default=100000, help='number of iterations to train')
parser.add_argument("--device", default=0, type=int, help="GPU id (default: 0)")
parser.add_argument("--lr", type=float, default=1e-3, help="Learning Rate. Default=4e-4")
parser.add_argument('--gamma', type=float, default=0.5, help='gamma')
parser.add_argument("--milestones", type=int, default=[10000,20000,60000], help="Sets the learning rate to the initial LR decayed by momentum every n epochs, Default: n=6")
parser.add_argument("--threads", type=int, default=4, help="Number of threads for data loader to use, Default: 1")

可是,训练到大概 9000 次迭代的时候 loss 变得巨大
8999it [32:18, 8.27it/s]Mar 29 23:46:06 iter---9000, loss_epoch---17572159662772990246912.000000, PSNR---5.812408

请问,您知道这是怎么回事嘛

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions